Evaluation of Mixed Deep Neural Networks for Reverberant Speech Enhancement

Area of Science:

• Artificial Intelligence
• Signal Processing
• Machine Learning

Background:

• Degraded speech signals in real-world environments pose challenges for voice recognition and analysis.
• Reverberation, caused by sound reflections, significantly degrades speech signal quality.
• Deep learning, particularly long short-term memory (LSTM) networks, shows promise for speech enhancement but suffers from high computational training costs.

Purpose of the Study:

• To evaluate hybrid neural network models for learning diverse reverberation conditions without prior information.
• To assess the effectiveness of combining LSTM and perceptron layers for speech signal enhancement.
• To compare the performance of hybrid models against pure LSTM networks in terms of quality, training time, and efficiency.

Main Methods:

• Trained and compared 120 artificial neural networks of eight different types, focusing on hybrid LSTM-perceptron architectures.
• Evaluated network performance using signal spectrum quality measurements and statistical validation.
• Measured and compared the training time of various network configurations.

Main Results:

• Certain hybrid LSTM-perceptron combinations achieved comparable results to pure LSTM networks with a fixed number of layers.
• Hybrid models demonstrated a significant reduction in training time, approximately 30%, compared to traditional LSTM approaches.
• The proposed hybrid networks offer efficiency gains without a substantial decrease in speech signal enhancement quality.

Conclusions:

• Hybrid neural networks are a viable and effective solution for speech signal enhancement in reverberant conditions.
• These models offer a practical alternative to pure LSTM networks, mitigating high computational costs and long training durations.
• The findings support the use of hybrid architectures for improving the robustness and efficiency of voice processing systems.